agent-ready-repo

A loop your coding agent can't cut corners in.
Loop engineering for any repo, any agent — a plan → execute → verify → review loop with gates it can't pass on red and a reviewer that reads every diff cold.

"You shouldn't be prompting coding agents anymore. You should be designing loops that prompt your agents." — Peter Steinberger

The leverage in agent coding has moved from the prompt to the loop — the system that plans, executes, verifies, and decides what's next without you babysitting each turn. The catch is the part everyone skips: a loop running unattended is also a loop making mistakes unattended. So the loop has to verify its own work harder than you would.

The core pack is that loop — built so the agent cannot self-certify its way out of it, and shipped as Markdown you can read, edit, and git diff. Not a framework, not a service. Files.

pip install agentbundle
agentbundle install --pack core git+https://github.com/eugenelim/agent-ready-repo

One line lands the loop in your repo. Any agent that reads a skill file — Claude Code, Codex, Cursor, Copilot, Kiro — inherits it.

The loop

Most agent workflows are prompt → code → ship. core replaces that one-shot path with a loop that separates the maker from the verifier and refuses to skip a step:

1. Plan with surfaced assumptions. Before any code, the agent writes what it's building, what it's not touching, and the assumptions it's making. When an assumption breaks mid-flight, the loop tells it to stop and surface — not guess past it.
2. Mechanical gates between every "done" and the actual finish. Lint, typecheck, and tests run as hard gates. There is no path through the loop that lets the agent declare success past a red gate.
3. Adversarial review in a fresh session. Once gates pass, a specialist reviewer reads the diff cold — no investment in the design, no sunk cost in the code. The loop iterates (fix → re-review) until the reviewer reports Clean — ready to commit.
4. Capture-learnings, because the model forgets and the repo doesn't. When a run trips over an undocumented convention, the gap lands as a candidate edit to CONVENTIONS.md. The agent's mistakes become the project's memory instead of evaporating at the end of the session.

The reviewer isn't one generic critic — three sharp lenses, dispatched by what the change actually touches:

The security lens shifts left: on security-boundary work it runs at spec stage as design guidance — catching a missing control as a one-sentence acceptance criterion instead of a post-implementation round-trip — and pulls its depth from a progressive-disclosure checklist library scoped to the boundaries a change actually crosses, so the review stays current (OWASP 2025, ASVS, CWE Top 25) without bloating the prompt. The quality lens holds every diff to a default quality floor — the universal maintainability smells and the mutation-testing mindset a strict static-analysis gate would enforce — applied by doctrine, whether or not such a gate is wired in.

A fourth subagent, implementer, is the loop's own executor for running independent tasks in parallel.

→ The full walkthrough — how the parts compose, and how it compares to vibe-coding, GitHub's Spec Kit, and Kiro's spec mode — is in The core pack as a system.

The catalogue

core is the loop. Everything else is à la carte — install only what your repo needs, at repo or user scope, one line each.

Repo-scope packs install into the current repo and build on core. User-scope packs install into ~/.claude/ (or your harness's home root) and follow you across every project. Swap core for any pack name in the install command above.

An ecosystem building block

The work-loop skill is a Markdown file. Each reviewer is a Markdown file. The bundler installs them as plain-text primitives — no SDK, no runtime, no service to host, no proprietary harness to lock into. That's what makes this a building block and not a walled garden:

• Harness-agnostic. One adapter pipeline projects the same primitives into Claude Code, Codex, Copilot, and Kiro layouts. Cursor reads AGENTS.md directly.
• Inspectable and forkable. The mechanics are prose you can read and git diff. When you outgrow a default, you edit the file — you don't file a feature request.
• Composable. Packs layer cleanly; your edits are never silently overwritten — colliding files land as *.upstream companions to merge, not clobber.

Loop engineering relocates judgment rather than removing it. Plain-text primitives keep that judgment where you can exercise it — build the loop like someone who intends to stay the engineer, not just the person who presses go.

Going deeper

Skills follow the agentskills.io specification — each is a self-contained folder with closed frontmatter and no hidden cross-skill dependencies, so they install, copy, and audit cleanly.

Contributing

Adding a pack, skill, or subagent? See CONTRIBUTING.md for the three contribution lanes, the pack source-of-truth split, and the gates your PR has to pass.

License

Licensed under either of Apache 2.0 or MIT at your option. Contributions are dual-licensed under the same terms unless you state otherwise.